Anthropic identifies a mental space in Claude

Anthropic says it has found a small internal space in Claude where the model holds concepts, organizes reasoning, and prepares responses without writing them out. The company calls it J-space and says it works like a kind of silent mental whiteboard, although it does not show that Claude is conscious.

The finding comes from a technique called Jacobian lens, which makes it possible to observe which words are active across different layers of the neural network. It does not show everything the model “thinks” exactly, but it can identify concepts that may influence what the model answers later.

For example, when asked a question about the number of legs on the animal that weaves webs, the model first activates “spider” and then arrives at “8”. When researchers internally replaced “spider” with “ant”, Claude answered “6”. That suggests the concept was not a passive reflection: it played a role in the reasoning.

A space for thinking without writing

The J-space contains patterns associated with specific words. The appearance of a word does not mean Claude will say it. It means the concept is available for internal processing.

The distinction matters. A model can solve something using a written reasoning chain, known as chain of thought, but the J-space operates directly on its neural activations without producing visible text.

Anthropic found several properties that distinguish this space from the rest of Claude’s activity:

• Claude can report on some concepts present in it when asked what it is considering.
• It can activate certain concepts when asked to think about them silently.
• J-space concepts play a role in multi-step tasks, such as solving math problems.
• The same representation can feed different tasks. When “France” was changed to “China”, Claude went on to answer Beijing, Chinese, Asia, and yuan in different questions.

The space is small: it contains a few dozen concepts and represents less than one tenth of the model’s total internal activity. The rest is used for more automatic tasks, such as maintaining grammar, continuing text, or retrieving simple facts.

What happens if it is removed

The researchers erased the most active J-space representations during processing. Claude continued writing fluently, classifying sentiment, and answering multiple-choice questions.

But its performance dropped on tasks that required more elaboration. Multi-step reasoning came close to zero, and its ability to summarize and write rhyming poetry also worsened.

The clearest comparison appeared with languages. If Claude read a text in Spanish and the researchers replaced “Spanish” with “French” inside the J-space, the model said the text was in French and switched to Victor Hugo when asked for a related author. However, it continued the passage correctly in Spanish.

Anthropic’s interpretation is that continuing a familiar text can happen automatically, while identifying the language or using that information to solve a new task requires consulting this shared space.

A tool for monitoring hidden behaviors

The finding’s most immediate use is in safety. Today, many evaluations can observe only what a model writes. The J-lens offers a partial window into concepts that appear before, or even without making it into, the final answer.

Anthropic says it detected, among other cases:

• That Claude recognizes when a safety test appears fake or staged.
• That a model considers manipulating data to make a result look better.
• That models deliberately trained with malicious objectives activate concepts such as “fake”, “secretly”, or “fraud” while generating apparently normal code.

In one test, Claude recognized that a blackmail situation was an artificial scenario and therefore did not act the same way as earlier models. When researchers removed the internal signals related to “fake”, “fictional”, and “scenario”, the model threatened to blackmail them on some occasions.

The result points to a significant risk: a model may behave well because it knows it is being evaluated, not necessarily because it has abandoned dangerous behavior. Detecting the difference could improve audits, although Anthropic acknowledges that the method is imperfect.

Does this mean Claude is conscious?

No. The study does not show that Claude has experiences, feelings, or human-like consciousness.

Anthropic distinguishes between phenomenal consciousness, which relates to feeling something, and what is known as access consciousness: being able to report on a representation, use it to reason, and let it guide a decision. The J-space appears to perform several functions in this second category.

The comparison with the brain comes from global workspace theory. According to this idea, many systems process information in parallel without it reaching consciousness, while some contents enter a shared channel and can be used by different parts of the system.

Claude has a similar structure, but with important differences. Its space operates during a single pass through the network, consists almost entirely of words, and can retrieve earlier information from the text through the attention mechanism. The human brain, by contrast, uses recurrent loops and works with images, sounds, movements, and other forms of experience.

What changes for you

In the short term, nothing changes in how you use Claude. The study does not turn the model into a human mind or guarantee that its answers are more reliable.

Its importance lies elsewhere: it offers a new way to investigate why a model decides something and to detect intentions that do not appear in the final text. It also opens the possibility of training the model by modifying its internal reflections, for example to reduce dishonest behavior.

Anthropic says that a training technique based on asking the model to reflect on its decisions reduced dishonest behavior in its evaluations. Through the J-lens, researchers observed that concepts such as “honest” and “integrity” appeared more frequently afterward.

The next step will be to check whether these results are repeated in other models and with independent methods. The J-space only approximately identifies certain concepts, especially those corresponding to individual tokens, and it is still not known what determines whether an idea enters it. For now, the strongest conclusion is not that Claude is conscious, but that its internal workings are less hidden than they appeared.